Agent-002: The LLM Call

🇬🇧 English

🇫🇷 Français

Agent-001 built the read-act-repeat loop with the simplest possible action: echo the line back. This stage keeps that loop completely intact and swaps in the only thing that changes between an echo and an agent – the action itself is now a call an LLM, for instance Claude AI.

The code

agents/agent-002-llm-call/agent.py looks almost identical to agent-001’s loop:

def run(input_stream, output_stream, client=None):
    client = client or Anthropic()

    while True:
        output_stream.write("User> ")
        output_stream.flush()
        line = input_stream.readline()
        if not line:
            break
        line = line.rstrip("\n")
        if not line:
            break
        message = client.messages.create(
            model=MODEL,
            max_tokens=1024,
            messages=[{"role": "user", "content": line}],
        )
        output_stream.write(f"Agent> {message.content[0].text}\n")

• Observation: still a line read from input_stream.
• Action: a messages.create call to Claude instead of an echo.
• Loop: unchanged from agent-001 – same prompt, same EOF/empty-line termination.

Every line is sent on its own, with no memory of earlier lines in the conversation. There’s no message history to maintain yet, so the agent treats each line as a fresh, independent prompt – the same way agent-001 treated each line as an independent echo.

Why client is a parameter

client defaults to a real Anthropic() instance (which reads ANTHROPIC_API_KEY from the environment), but it can be overridden. That’s the same idea as passing in input_stream/output_stream instead of hardcoding sys.stdin/sys.stdout: it lets test_agent.py substitute a fake client that returns a canned reply, so the test suite runs without hitting the network or needing an API key.

Running it

export ANTHROPIC_API_KEY=<your key>
python agents/agent-002-llm-call/agent.py

Type a line, press enter, see Claude’s reply. Press enter on an empty line to stop.

What’s next

This stage is stateless – Agent has no idea what you said on the previous line. Agent-003 starts carrying conversation history between turns, and later stages will add tools so Agent can act on more than just text.